llmfind llms should run on premises

y intelligence indexx $ per 1M tokens, 600 in : 5 out, logsize output tokens/s ● api● open weights, priced by card write-off◌ new this fortnight
0204060 $0.001$0.01$0.1$1$10$100 $ / 1M tokens → ↑ intelligence index Claude Opus 5.5 · anthropic intelligence index 59.1 · $4.13 / 1M · 95 tok/sClaude Opus 5.5 GPT-6 Astra · openai intelligence index 56.9 · $10.3 / 1M · 49 tok/s GPT-6.1 Sol · openai intelligence index 54.6 · $2.07 / 1M · 67 tok/sGPT-6.1 Sol Claude Sonnet 5.5 · anthropic intelligence index 54.5 · $2.07 / 1M · 97 tok/s GPT-6.1 Terra · openai intelligence index 51.3 · $1.03 / 1M · 78 tok/sGPT-6.1 Terra Gemini 4 Argon · google intelligence index 50.4 · $2.07 / 1M · 91 tok/s Claude Fable 5.1 · anthropic intelligence index 50 · $10.3 / 1M · 53 tok/s Kimi K3.1 Pro · moonshotai · open weights intelligence index 46.5 · $2.07 / 1M · 27 tok/s GPT-6 Sol · openai intelligence index 44.2 · $2.07 / 1M · 129 tok/s Claude Opus 5 · anthropic intelligence index 44.2 · $5.17 / 1M · 49 tok/s Muse Wonder · meta intelligence index 43.6 · $4.13 / 1M · 164 tok/s Muse Spark 1.3 · meta intelligence index 36.6 · $1.27 / 1M · 146 tok/s Claude Haiku 5.5 · anthropic intelligence index 36.5 · $0.103 / 1M · 165 tok/sClaude Haiku 5.5 Grok 4.7 · x-ai intelligence index 35.5 · $2.03 / 1M · 55 tok/s Hy4 · tencent · open weights intelligence index 34.6 · $0.424 / 1M Qwen3.8 Max (0902) · qwen intelligence index 33.7 · $2.03 / 1M GPT-5.6 Terra · openai intelligence index 32.3 · $2.08 / 1M · 95 tok/s GLM 5.3 · z-ai · open weights intelligence index 31.7 · $0.0661 / 1M on 6× B200 · 72 tok/sGLM 5.3 Grok 4.6 · x-ai intelligence index 31 · $2.03 / 1M · 63 tok/s Gemini 3.8 Flash · google intelligence index 30.8 · $0.852 / 1M · 136 tok/s MiMo-V2.6-Pro · xiaomi · open weights intelligence index 29.9 · $0.0599 / 1M on 7× B200 · 41 tok/sMiMo-V2.6-Pro MiniMax M3.1 Flash · minimax · open weights intelligence index 29.1 · $0.0194 / 1M on 2× B200 · 152 tok/sMiniMax M3.1 Flash DeepSeek V4.1 Flash 1017 · deepseek · open weights intelligence index 28.6 · $0.0335 / 1M on 6× B200 · 155 tok/s Step 5 Preview · stepfun · open weights intelligence index 28.5 · $0.0386 / 1M on 5× B200 · 150 tok/s Solar Pro 4 · upstage intelligence index 28.2 · $0.154 / 1M · 58 tok/s GPT-6 Luna · openai intelligence index 27.9 · $0.103 / 1M · 118 tok/s GLM 5.3 Flash · z-ai · open weights intelligence index 27.8 · $0.0251 / 1M on 3× B200 · 55 tok/s MiMo-V2.6-Flash · xiaomi · open weights intelligence index 26.7 · $0.0273 / 1M on 3× B200 · 37 tok/s Mistral Large 4 · mistralai · open weights intelligence index 26.1 · $0.0974 / 1M on 7× B200 · 128 tok/s Muse Spark Open · meta · open weights intelligence index 24.9 · $0.0787 / 1M on 6× B200 · 214 tok/s GPT-5.6 Luna · openai intelligence index 23.8 · $0.208 / 1M · 105 tok/s DeepSeek V4 Pro 0813 · deepseek · open weights intelligence index 23.1 · $0.154 / 1M · 79 tok/s Qwen3.8 27B · qwen · open weights intelligence index 20.1 · $0.052 / 1M on 1× B200 · 36 tok/s MiniMax M3 · minimax · open weights intelligence index 18.9 · $0.0375 / 1M on 3× B200 · 121 tok/s Beam 1.0 501B A23B · reflection · open weights intelligence index 17.7 · $0.0431 / 1M on 4× B200 Kimi K3 · moonshotai · open weights intelligence index 17.4 · $0.323 / 1M · 50 tok/s Nemotron 3.5 Super VL · nvidia · open weights intelligence index 13.6 · $0.153 / 1M · 266 tok/s Inkling Small · thinkingmachines · open weights intelligence index 12.9 · $0.0175 / 1M on 2× B200 · 81 tok/sInkling Small DeepSeek V4 Pro 0423 · deepseek · open weights intelligence index 12.6 · $0.714 / 1M Inkling · thinkingmachines · open weights intelligence index 12.5 · $0.0627 / 1M on 7× B200 · 139 tok/s Nemotron 3 Ultra · nvidia · open weights intelligence index 11.8 · $0.103 / 1M on 4× B200 Gemini 3.5 Flash Lite · google intelligence index 11.7 · $0.318 / 1M · 392 tok/s Claude Haiku 4.5 · anthropic intelligence index 7.4 · $1.03 / 1M Muse Glimmer 30B · meta · open weights intelligence index 6.3 · $0.0558 / 1M on 1× B200 Command A+ · cohere intelligence index 6.2 · $0.31 / 1M · 247 tok/s Gemini 3.1 Flash Lite Preview · google intelligence index 6.2 · $0.26 / 1M Nemotron 3.5 Lightning · nvidia · open weights intelligence index 5.5 · $0.00562 / 1M on 1× B200 · 293 tok/sNemotron 3.5 Lightning Granite 4.2 8B · ibm-granite · open weights intelligence index 5.1 · $0.0165 / 1M on 1× B200 · 124 tok/s Mistral Medium 3.5 · mistralai intelligence index 4.6 · $1.55 / 1M · 172 tok/s Gemma 4 31B · google · open weights intelligence index 4.4 · $0.0586 / 1M on 1× B200 · 35 tok/s Mistral Small 4 · mistralai · open weights intelligence index 4.3 · $0.0105 / 1M on 1× B200 · 160 tok/s North Mini Code · cohere · open weights intelligence index 3.4 · $0.00535 / 1M on 1× B200 · 99 tok/sNorth Mini Code

try it — click a dot: the model opens on arena.ai (or OpenRouter chat when arena does not list it) and this prompt lands in your clipboard, paste it there:

Classify this support ticket as billing, bug, feature or other. Reply with one word only.

Ticket: Since yesterday's update the export button does nothing.
draw a dotclick the chart to fill in accuracy and cost. dots live in a cookie - only you see them.
new model?
modelintelligence$ / 1Mtok/sruns onsince
Claude Opus 5.5 anthropic●59.14.1395 2026-09try
GPT-6 Astra openai●56.910.349 2026-09try
GPT-6.1 Sol openai●54.62.0767 2026-09try
Claude Sonnet 5.5 anthropic●54.52.0797 2026-09try
GPT-6.1 Terra openai●51.31.0378 2026-10try
Gemini 4 Argon google●50.42.0791 2026-10try
Claude Fable 5.1 anthropic●5010.353 2026-09try
Kimi K3.1 Pro moonshotai●46.52.0727 api price2026-10try
GPT-6 Sol openai●44.22.07129 2026-09try
Claude Opus 5 anthropic●44.25.1749 2026-07try
Muse Wonder meta●43.64.13164 2026-10try
Muse Spark 1.3 meta●36.61.27146 2026-09try
Claude Haiku 5.5 anthropic●36.50.103165 2026-10try
Grok 4.7 x-ai●35.52.0355 2026-09try
Hy4 tencent●34.60.424 api price2026-11try
Qwen3.8 Max (0902) qwen●33.72.03 2026-09try
GPT-5.6 Terra openai●32.32.0895 2026-07try
GLM 5.3 z-ai●31.70.066172 6× B200 753B, 35B active2026-08try
Grok 4.6 x-ai●312.0363 2026-08try
Gemini 3.8 Flash google●30.80.852136 2026-09try
MiMo-V2.6-Pro xiaomi●29.90.059941 7× B200 1024B, 32B active2026-09try
MiniMax M3.1 Flash minimax●29.10.0194152 2× B200 220B, 10B active2026-10try
DeepSeek V4.1 Flash 1017 deepseek●28.60.0335155 6× B200 763B, 18B active2026-10try
Step 5 Preview stepfun●28.50.0386150 5× B200 604B, 21B active2026-10try
Solar Pro 4 upstage●28.20.15458 2026-08try
GPT-6 Luna openai●27.90.103118 2026-09try
GLM 5.3 Flash z-ai●27.80.025155 3× B200 321B, 13B active2026-08try
MiMo-V2.6-Flash xiaomi●26.70.027337 3× B200 311B, 15B active2026-09try
Mistral Large 4 mistralai●26.10.0974128 7× B200 1000B, 52B active2026-10try
Muse Spark Open meta●24.90.0787214 6× B200 816B, 42B active2026-10try
GPT-5.6 Luna openai●23.80.208105 2026-07try
DeepSeek V4 Pro 0813 deepseek●23.10.15479 api price2026-08try
Qwen3.8 27B qwen●20.10.05236 1× B200 28B2026-08try
MiniMax M3 minimax●18.90.0375121 3× B200 427B, 20B active2026-05try
Beam 1.0 501B A23B reflection●17.70.0431 4× B200 501B, 23B active2026-10try
Kimi K3 moonshotai●17.40.32350 api price2026-07try
Nemotron 3.5 Super VL nvidia●13.60.153266 api price2026-10try
Inkling Small thinkingmachines●12.90.017581 2× B200 266B, 9B active2026-07try
DeepSeek V4 Pro 0423 deepseek●12.60.714 api price2026-04try
Inkling thinkingmachines●12.50.0627139 7× B200 952B, 33B active2026-07try
Nemotron 3 Ultra nvidia●11.80.103 4× B200 550B, 55B active2026-06try
Gemini 3.5 Flash Lite google●11.70.318392 2026-07try
Claude Haiku 4.5 anthropic●7.41.03 2025-10try
Muse Glimmer 30B meta●6.30.0558 1× B200 30B2026-08try
Command A+ cohere●6.20.31247 2026-09try
Gemini 3.1 Flash Lite Preview google●6.20.26 2026-03try
Nemotron 3.5 Lightning nvidia●5.50.00562293 1× B200 30B, 3B active2026-08try
Granite 4.2 8B ibm-granite●5.10.0165124 1× B200 9B2026-08try
Mistral Medium 3.5 mistralai●4.61.55172 2026-04try
Gemma 4 31B google●4.40.058635 1× B200 31B2026-04try
Mistral Small 4 mistralai●4.30.0105160 1× B200 119B, 6B active2026-03try
North Mini Code cohere●3.40.0053599 1× B200 30B, 3B active2026-06try
how open weights are priced

A card is written off over 3 years at 50% utilization, plus electricity at $0.2/kWh, times 1.3 for the box around it. The cheapest card (at most 8 per box) whose memory holds the fp8 weights wins. An output token costs card-hours ÷ (peak TFLOPS × 0.2 ÷ 2 × active parameters), an input token the same at 0.6. No margin, no idle box, no engineer: a floor, not a quote. Parameter counts come from Hugging Face; for mixture-of-experts models the active share is estimated.

card$GBTFLOPSW$ / h
RTX 50902000322105750.50
H10030000809897003.33
B20040000180225010004.48